src/glsl/nir/nir.c

   1 /*
   2  * Copyright © 2014 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *
  26  */
  27
  28 #include "nir.h"
  29 #include <assert.h>
  30
  31 nir_shader *
  32 nir_shader_create(void *mem_ctx)
  33 {
  34    nir_shader *shader = ralloc(mem_ctx, nir_shader);
  35
  36    shader->uniforms = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  37                                               _mesa_key_string_equal);
  38    shader->inputs = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  39                                             _mesa_key_string_equal);
  40    shader->outputs = _mesa_hash_table_create(shader, _mesa_key_hash_string,
  41                                              _mesa_key_string_equal);
  42
  43    shader->num_user_structures = 0;
  44    shader->user_structures = NULL;
  45
  46    exec_list_make_empty(&shader->functions);
  47    exec_list_make_empty(&shader->registers);
  48    exec_list_make_empty(&shader->globals);
  49    exec_list_make_empty(&shader->system_values);
  50    shader->reg_alloc = 0;
  51
  52    shader->num_inputs = 0;
  53    shader->num_outputs = 0;
  54    shader->num_uniforms = 0;
  55
  56    return shader;
  57 }
  58
  59 static nir_register *
  60 reg_create(void *mem_ctx, struct exec_list *list)
  61 {
  62    nir_register *reg = ralloc(mem_ctx, nir_register);
  63
  64    reg->uses = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
  65    reg->defs = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
  66    reg->if_uses = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
  67
  68    reg->num_components = 0;
  69    reg->num_array_elems = 0;
  70    reg->is_packed = false;
  71    reg->name = NULL;
  72
  73    exec_list_push_tail(list, &reg->node);
  74
  75    return reg;
  76 }
  77
  78 nir_register *
  79 nir_global_reg_create(nir_shader *shader)
  80 {
  81    nir_register *reg = reg_create(shader, &shader->registers);
  82    reg->index = shader->reg_alloc++;
  83    reg->is_global = true;
  84
  85    return reg;
  86 }
  87
  88 nir_register *
  89 nir_local_reg_create(nir_function_impl *impl)
  90 {
  91    nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers);
  92    reg->index = impl->reg_alloc++;
  93    reg->is_global = false;
  94
  95    return reg;
  96 }
  97
  98 void
  99 nir_reg_remove(nir_register *reg)
 100 {
 101    exec_node_remove(&reg->node);
 102 }
 103
 104 nir_function *
 105 nir_function_create(nir_shader *shader, const char *name)
 106 {
 107    nir_function *func = ralloc(shader, nir_function);
 108
 109    exec_list_push_tail(&shader->functions, &func->node);
 110    exec_list_make_empty(&func->overload_list);
 111    func->name = name;
 112    func->shader = shader;
 113
 114    return func;
 115 }
 116
 117 nir_function_overload *
 118 nir_function_overload_create(nir_function *func)
 119 {
 120    void *mem_ctx = ralloc_parent(func);
 121
 122    nir_function_overload *overload = ralloc(mem_ctx, nir_function_overload);
 123
 124    overload->num_params = 0;
 125    overload->params = NULL;
 126    overload->return_type = glsl_void_type();
 127    overload->impl = NULL;
 128
 129    exec_list_push_tail(&func->overload_list, &overload->node);
 130    overload->function = func;
 131
 132    return overload;
 133 }
 134
 135 nir_src nir_src_copy(nir_src src, void *mem_ctx)
 136 {
 137    nir_src ret;
 138    ret.is_ssa = src.is_ssa;
 139    if (ret.is_ssa) {
 140       ret.ssa = src.ssa;
 141    } else {
 142       ret.reg.base_offset = src.reg.base_offset;
 143       ret.reg.reg = src.reg.reg;
 144       if (src.reg.indirect) {
 145          ret.reg.indirect = ralloc(mem_ctx, nir_src);
 146          *ret.reg.indirect = *src.reg.indirect;
 147       } else {
 148          ret.reg.indirect = NULL;
 149       }
 150    }
 151
 152    return ret;
 153 }
 154
 155 nir_dest nir_dest_copy(nir_dest dest, void *mem_ctx)
 156 {
 157    nir_dest ret;
 158    ret.is_ssa = dest.is_ssa;
 159    if (ret.is_ssa) {
 160       ret.ssa = dest.ssa;
 161    } else {
 162       ret.reg.base_offset = dest.reg.base_offset;
 163       ret.reg.reg = dest.reg.reg;
 164       if (dest.reg.indirect) {
 165          ret.reg.indirect = ralloc(mem_ctx, nir_src);
 166          *ret.reg.indirect = *dest.reg.indirect;
 167       } else {
 168          ret.reg.indirect = NULL;
 169       }
 170    }
 171
 172    return ret;
 173 }
 174
 175 static inline void
 176 block_add_pred(nir_block *block, nir_block *pred)
 177 {
 178    _mesa_set_add(block->predecessors, _mesa_hash_pointer(pred), pred);
 179 }
 180
 181 static void
 182 cf_init(nir_cf_node *node, nir_cf_node_type type)
 183 {
 184    exec_node_init(&node->node);
 185    node->parent = NULL;
 186    node->type = type;
 187 }
 188
 189 static void
 190 link_blocks(nir_block *pred, nir_block *succ1, nir_block *succ2)
 191 {
 192    pred->successors[0] = succ1;
 193    block_add_pred(succ1, pred);
 194
 195    pred->successors[1] = succ2;
 196    if (succ2 != NULL)
 197       block_add_pred(succ2, pred);
 198 }
 199
 200 static void
 201 unlink_blocks(nir_block *pred, nir_block *succ)
 202 {
 203    if (pred->successors[0] == succ) {
 204       pred->successors[0] = pred->successors[1];
 205       pred->successors[1] = NULL;
 206    } else {
 207       assert(pred->successors[1] == succ);
 208       pred->successors[1] = NULL;
 209    }
 210
 211    struct set_entry *entry = _mesa_set_search(succ->predecessors,
 212                                               _mesa_hash_pointer(pred), pred);
 213
 214    assert(entry);
 215
 216    _mesa_set_remove(succ->predecessors, entry);
 217 }
 218
 219 static void
 220 unlink_block_successors(nir_block *block)
 221 {
 222    if (block->successors[0] != NULL)
 223       unlink_blocks(block, block->successors[0]);
 224    if (block->successors[1] != NULL)
 225       unlink_blocks(block, block->successors[1]);
 226 }
 227
 228
 229 nir_function_impl *
 230 nir_function_impl_create(nir_function_overload *overload)
 231 {
 232    assert(overload->impl == NULL);
 233
 234    void *mem_ctx = ralloc_parent(overload);
 235
 236    nir_function_impl *impl = ralloc(mem_ctx, nir_function_impl);
 237
 238    overload->impl = impl;
 239    impl->overload = overload;
 240
 241    cf_init(&impl->cf_node, nir_cf_node_function);
 242
 243    exec_list_make_empty(&impl->body);
 244    exec_list_make_empty(&impl->registers);
 245    exec_list_make_empty(&impl->locals);
 246    impl->num_params = 0;
 247    impl->params = NULL;
 248    impl->return_var = NULL;
 249    impl->reg_alloc = 0;
 250    impl->ssa_alloc = 0;
 251    impl->valid_metadata = nir_metadata_none;
 252
 253    /* create start & end blocks */
 254    nir_block *start_block = nir_block_create(mem_ctx);
 255    nir_block *end_block = nir_block_create(mem_ctx);
 256    start_block->cf_node.parent = &impl->cf_node;
 257    end_block->cf_node.parent = &impl->cf_node;
 258    impl->start_block = start_block;
 259    impl->end_block = end_block;
 260
 261    exec_list_push_tail(&impl->body, &start_block->cf_node.node);
 262
 263    start_block->successors[0] = end_block;
 264    block_add_pred(end_block, start_block);
 265
 266    return impl;
 267 }
 268
 269 nir_block *
 270 nir_block_create(void *mem_ctx)
 271 {
 272    nir_block *block = ralloc(mem_ctx, nir_block);
 273
 274    cf_init(&block->cf_node, nir_cf_node_block);
 275
 276    block->successors[0] = block->successors[1] = NULL;
 277    block->predecessors = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
 278    block->imm_dom = NULL;
 279    block->dom_frontier = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
 280
 281    exec_list_make_empty(&block->instr_list);
 282
 283    return block;
 284 }
 285
 286 static inline void
 287 src_init(nir_src *src)
 288 {
 289    src->is_ssa = false;
 290    src->reg.reg = NULL;
 291    src->reg.indirect = NULL;
 292    src->reg.base_offset = 0;
 293 }
 294
 295 nir_if *
 296 nir_if_create(void *mem_ctx)
 297 {
 298    nir_if *if_stmt = ralloc(mem_ctx, nir_if);
 299
 300    cf_init(&if_stmt->cf_node, nir_cf_node_if);
 301    src_init(&if_stmt->condition);
 302
 303    nir_block *then = nir_block_create(mem_ctx);
 304    exec_list_make_empty(&if_stmt->then_list);
 305    exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node);
 306    then->cf_node.parent = &if_stmt->cf_node;
 307
 308    nir_block *else_stmt = nir_block_create(mem_ctx);
 309    exec_list_make_empty(&if_stmt->else_list);
 310    exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node);
 311    else_stmt->cf_node.parent = &if_stmt->cf_node;
 312
 313    return if_stmt;
 314 }
 315
 316 nir_loop *
 317 nir_loop_create(void *mem_ctx)
 318 {
 319    nir_loop *loop = ralloc(mem_ctx, nir_loop);
 320
 321    cf_init(&loop->cf_node, nir_cf_node_loop);
 322
 323    nir_block *body = nir_block_create(mem_ctx);
 324    exec_list_make_empty(&loop->body);
 325    exec_list_push_tail(&loop->body, &body->cf_node.node);
 326    body->cf_node.parent = &loop->cf_node;
 327
 328    body->successors[0] = body;
 329    block_add_pred(body, body);
 330
 331    return loop;
 332 }
 333
 334 static void
 335 instr_init(nir_instr *instr, nir_instr_type type)
 336 {
 337    instr->type = type;
 338    instr->block = NULL;
 339    exec_node_init(&instr->node);
 340 }
 341
 342 static void
 343 dest_init(nir_dest *dest)
 344 {
 345    dest->is_ssa = false;
 346    dest->reg.reg = NULL;
 347    dest->reg.indirect = NULL;
 348    dest->reg.base_offset = 0;
 349 }
 350
 351 static void
 352 alu_dest_init(nir_alu_dest *dest)
 353 {
 354    dest_init(&dest->dest);
 355    dest->saturate = false;
 356    dest->write_mask = 0xf;
 357 }
 358
 359 static void
 360 alu_src_init(nir_alu_src *src)
 361 {
 362    src_init(&src->src);
 363    src->abs = src->negate = false;
 364    src->swizzle[0] = 0;
 365    src->swizzle[1] = 1;
 366    src->swizzle[2] = 2;
 367    src->swizzle[3] = 3;
 368 }
 369
 370 nir_alu_instr *
 371 nir_alu_instr_create(void *mem_ctx, nir_op op)
 372 {
 373    unsigned num_srcs = nir_op_infos[op].num_inputs;
 374    nir_alu_instr *instr =
 375       ralloc_size(mem_ctx,
 376                   sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src));
 377
 378    instr_init(&instr->instr, nir_instr_type_alu);
 379    instr->op = op;
 380    alu_dest_init(&instr->dest);
 381    for (unsigned i = 0; i < num_srcs; i++)
 382       alu_src_init(&instr->src[i]);
 383
 384    instr->has_predicate = false;
 385    src_init(&instr->predicate);
 386
 387    return instr;
 388 }
 389
 390 nir_jump_instr *
 391 nir_jump_instr_create(void *mem_ctx, nir_jump_type type)
 392 {
 393    nir_jump_instr *instr = ralloc(mem_ctx, nir_jump_instr);
 394    instr_init(&instr->instr, nir_instr_type_jump);
 395    instr->type = type;
 396    return instr;
 397 }
 398
 399 nir_load_const_instr *
 400 nir_load_const_instr_create(void *mem_ctx)
 401 {
 402    nir_load_const_instr *instr = ralloc(mem_ctx, nir_load_const_instr);
 403    instr_init(&instr->instr, nir_instr_type_load_const);
 404
 405    dest_init(&instr->dest);
 406    instr->num_components = 0;
 407    instr->array_elems = 0;
 408
 409    instr->has_predicate = false;
 410    src_init(&instr->predicate);
 411
 412    return instr;
 413 }
 414
 415 nir_intrinsic_instr *
 416 nir_intrinsic_instr_create(void *mem_ctx, nir_intrinsic_op op)
 417 {
 418    unsigned num_srcs = nir_intrinsic_infos[op].num_srcs;
 419    nir_intrinsic_instr *instr =
 420       ralloc_size(mem_ctx,
 421                   sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src));
 422
 423    instr_init(&instr->instr, nir_instr_type_intrinsic);
 424    instr->intrinsic = op;
 425
 426    if (nir_intrinsic_infos[op].has_dest)
 427       dest_init(&instr->dest);
 428
 429    for (unsigned i = 0; i < num_srcs; i++)
 430       src_init(&instr->src[i]);
 431
 432    instr->has_predicate = false;
 433    src_init(&instr->predicate);
 434
 435    return instr;
 436 }
 437
 438 nir_call_instr *
 439 nir_call_instr_create(void *mem_ctx, nir_function_overload *callee)
 440 {
 441    nir_call_instr *instr = ralloc(mem_ctx, nir_call_instr);
 442    instr_init(&instr->instr, nir_instr_type_call);
 443
 444    instr->callee = callee;
 445    instr->num_params = callee->num_params;
 446    instr->params = ralloc_array(mem_ctx, nir_deref_var *, instr->num_params);
 447    instr->return_deref = NULL;
 448
 449    instr->has_predicate = false;
 450    src_init(&instr->predicate);
 451
 452    return instr;
 453 }
 454
 455 nir_tex_instr *
 456 nir_tex_instr_create(void *mem_ctx, unsigned num_srcs)
 457 {
 458    nir_tex_instr *instr = ralloc(mem_ctx, nir_tex_instr);
 459    instr_init(&instr->instr, nir_instr_type_texture);
 460
 461    dest_init(&instr->dest);
 462
 463    instr->num_srcs = num_srcs;
 464    for (unsigned i = 0; i < num_srcs; i++)
 465       src_init(&instr->src[i]);
 466
 467    instr->has_predicate = false;
 468    src_init(&instr->predicate);
 469
 470    return instr;
 471 }
 472
 473 nir_phi_instr *
 474 nir_phi_instr_create(void *mem_ctx)
 475 {
 476    nir_phi_instr *instr = ralloc(mem_ctx, nir_phi_instr);
 477    instr_init(&instr->instr, nir_instr_type_phi);
 478
 479    dest_init(&instr->dest);
 480    exec_list_make_empty(&instr->srcs);
 481    return instr;
 482 }
 483
 484 nir_parallel_copy_instr *
 485 nir_parallel_copy_instr_create(void *mem_ctx)
 486 {
 487    nir_parallel_copy_instr *instr = ralloc(mem_ctx, nir_parallel_copy_instr);
 488    instr_init(&instr->instr, nir_instr_type_parallel_copy);
 489
 490    instr->at_end = false;
 491    exec_list_make_empty(&instr->copies);
 492
 493    return instr;
 494 }
 495
 496 nir_ssa_undef_instr *
 497 nir_ssa_undef_instr_create(void *mem_ctx)
 498 {
 499    nir_ssa_undef_instr *instr = ralloc(mem_ctx, nir_ssa_undef_instr);
 500    instr_init(&instr->instr, nir_instr_type_ssa_undef);
 501
 502    instr->def.name = NULL;
 503    instr->def.parent_instr = &instr->instr;
 504
 505    return instr;
 506 }
 507
 508 nir_deref_var *
 509 nir_deref_var_create(void *mem_ctx, nir_variable *var)
 510 {
 511    nir_deref_var *deref = ralloc(mem_ctx, nir_deref_var);
 512    deref->deref.deref_type = nir_deref_type_var;
 513    deref->deref.child = NULL;
 514    deref->deref.type = var->type;
 515    deref->var = var;
 516    return deref;
 517 }
 518
 519 nir_deref_array *
 520 nir_deref_array_create(void *mem_ctx)
 521 {
 522    nir_deref_array *deref = ralloc(mem_ctx, nir_deref_array);
 523    deref->deref.deref_type = nir_deref_type_array;
 524    deref->deref.child = NULL;
 525    deref->has_indirect = false;
 526    src_init(&deref->indirect);
 527    deref->base_offset = 0;
 528    return deref;
 529 }
 530
 531 nir_deref_struct *
 532 nir_deref_struct_create(void *mem_ctx, const char *field)
 533 {
 534    nir_deref_struct *deref = ralloc(mem_ctx, nir_deref_struct);
 535    deref->deref.deref_type = nir_deref_type_struct;
 536    deref->deref.child = NULL;
 537    deref->elem = ralloc_strdup(deref, field);
 538    return deref;
 539 }
 540
 541 static nir_deref_var *
 542 copy_deref_var(void *mem_ctx, nir_deref_var *deref)
 543 {
 544    nir_deref_var *ret = nir_deref_var_create(mem_ctx, deref->var);
 545    if (deref->deref.child)
 546       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 547    return ret;
 548 }
 549
 550 static nir_deref_array *
 551 copy_deref_array(void *mem_ctx, nir_deref_array *deref)
 552 {
 553    nir_deref_array *ret = nir_deref_array_create(mem_ctx);
 554    ret->base_offset = deref->base_offset;
 555    if (deref->has_indirect)
 556       ret->indirect = deref->indirect;
 557    ret->deref.type = deref->deref.type;
 558    if (deref->deref.child)
 559       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 560    return ret;
 561 }
 562
 563 static nir_deref_struct *
 564 copy_deref_struct(void *mem_ctx, nir_deref_struct *deref)
 565 {
 566    nir_deref_struct *ret = nir_deref_struct_create(mem_ctx, deref->elem);
 567    ret->deref.type = deref->deref.type;
 568    if (deref->deref.child)
 569       ret->deref.child = nir_copy_deref(mem_ctx, deref->deref.child);
 570    return ret;
 571 }
 572
 573 nir_deref *
 574 nir_copy_deref(void *mem_ctx, nir_deref *deref)
 575 {
 576    switch (deref->deref_type) {
 577    case nir_deref_type_var:
 578       return &copy_deref_var(mem_ctx, nir_deref_as_var(deref))->deref;
 579    case nir_deref_type_array:
 580       return &copy_deref_array(mem_ctx, nir_deref_as_array(deref))->deref;
 581    case nir_deref_type_struct:
 582       return &copy_deref_struct(mem_ctx, nir_deref_as_struct(deref))->deref;
 583    default:
 584       unreachable("Invalid dereference type");
 585    }
 586
 587    return NULL;
 588 }
 589
 590
 591 /**
 592  * \name Control flow modification
 593  *
 594  * These functions modify the control flow tree while keeping the control flow
 595  * graph up-to-date. The invariants respected are:
 596  * 1. Each then statement, else statement, or loop body must have at least one
 597  *    control flow node.
 598  * 2. Each if-statement and loop must have one basic block before it and one
 599  *    after.
 600  * 3. Two basic blocks cannot be directly next to each other.
 601  * 4. If a basic block has a jump instruction, there must be only one and it
 602  *    must be at the end of the block.
 603  * 5. The CFG must always be connected - this means that we must insert a fake
 604  *    CFG edge for loops with no break statement.
 605  *
 606  * The purpose of the second one is so that we have places to insert code during
 607  * GCM, as well as eliminating the possibility of critical edges.
 608  */
 609 /*@{*/
 610
 611 static void
 612 link_non_block_to_block(nir_cf_node *node, nir_block *block)
 613 {
 614    if (node->type == nir_cf_node_if) {
 615       /*
 616        * We're trying to link an if to a block after it; this just means linking
 617        * the last block of the then and else branches.
 618        */
 619
 620       nir_if *if_stmt = nir_cf_node_as_if(node);
 621
 622       nir_cf_node *last_then = nir_if_last_then_node(if_stmt);
 623       assert(last_then->type == nir_cf_node_block);
 624       nir_block *last_then_block = nir_cf_node_as_block(last_then);
 625
 626       nir_cf_node *last_else = nir_if_last_else_node(if_stmt);
 627       assert(last_else->type == nir_cf_node_block);
 628       nir_block *last_else_block = nir_cf_node_as_block(last_else);
 629
 630       if (exec_list_is_empty(&last_then_block->instr_list) ||
 631           nir_block_last_instr(last_then_block)->type != nir_instr_type_jump) {
 632          unlink_block_successors(last_then_block);
 633          link_blocks(last_then_block, block, NULL);
 634       }
 635
 636       if (exec_list_is_empty(&last_else_block->instr_list) ||
 637           nir_block_last_instr(last_else_block)->type != nir_instr_type_jump) {
 638          unlink_block_successors(last_else_block);
 639          link_blocks(last_else_block, block, NULL);
 640       }
 641    } else {
 642       assert(node->type == nir_cf_node_loop);
 643
 644       /*
 645        * We can only get to this codepath if we're inserting a new loop, or
 646        * at least a loop with no break statements; we can't insert break
 647        * statements into a loop when we haven't inserted it into the CFG
 648        * because we wouldn't know which block comes after the loop
 649        * and therefore, which block should be the successor of the block with
 650        * the break). Therefore, we need to insert a fake edge (see invariant
 651        * #5).
 652        */
 653
 654       nir_loop *loop = nir_cf_node_as_loop(node);
 655
 656       nir_cf_node *last = nir_loop_last_cf_node(loop);
 657       assert(last->type == nir_cf_node_block);
 658       nir_block *last_block =  nir_cf_node_as_block(last);
 659
 660       last_block->successors[1] = block;
 661       block_add_pred(block, last_block);
 662    }
 663 }
 664
 665 static void
 666 link_block_to_non_block(nir_block *block, nir_cf_node *node)
 667 {
 668    if (node->type == nir_cf_node_if) {
 669       /*
 670        * We're trying to link a block to an if after it; this just means linking
 671        * the block to the first block of the then and else branches.
 672        */
 673
 674       nir_if *if_stmt = nir_cf_node_as_if(node);
 675
 676       nir_cf_node *first_then = nir_if_first_then_node(if_stmt);
 677       assert(first_then->type == nir_cf_node_block);
 678       nir_block *first_then_block = nir_cf_node_as_block(first_then);
 679
 680       nir_cf_node *first_else = nir_if_first_else_node(if_stmt);
 681       assert(first_else->type == nir_cf_node_block);
 682       nir_block *first_else_block = nir_cf_node_as_block(first_else);
 683
 684       unlink_block_successors(block);
 685       link_blocks(block, first_then_block, first_else_block);
 686    } else {
 687       /*
 688        * For similar reasons as the corresponding case in
 689        * link_non_block_to_block(), don't worry about if the loop header has
 690        * any predecessors that need to be unlinked.
 691        */
 692
 693       assert(node->type == nir_cf_node_loop);
 694
 695       nir_loop *loop = nir_cf_node_as_loop(node);
 696
 697       nir_cf_node *loop_header = nir_loop_first_cf_node(loop);
 698       assert(loop_header->type == nir_cf_node_block);
 699       nir_block *loop_header_block = nir_cf_node_as_block(loop_header);
 700
 701       unlink_block_successors(block);
 702       link_blocks(block, loop_header_block, NULL);
 703    }
 704
 705 }
 706
 707 /**
 708  * Takes a basic block and inserts a new empty basic block before it, making its
 709  * predecessors point to the new block. This essentially splits the block into
 710  * an empty header and a body so that another non-block CF node can be inserted
 711  * between the two. Note that this does *not* link the two basic blocks, so
 712  * some kind of cleanup *must* be performed after this call.
 713  */
 714
 715 static nir_block *
 716 split_block_beginning(nir_block *block)
 717 {
 718    nir_block *new_block = nir_block_create(ralloc_parent(block));
 719    new_block->cf_node.parent = block->cf_node.parent;
 720    exec_node_insert_node_before(&block->cf_node.node, &new_block->cf_node.node);
 721
 722    struct set_entry *entry;
 723    set_foreach(block->predecessors, entry) {
 724       nir_block *pred = (nir_block *) entry->key;
 725
 726       unlink_blocks(pred, block);
 727       link_blocks(pred, new_block, NULL);
 728    }
 729
 730    return new_block;
 731 }
 732
 733 /**
 734  * Moves the successors of source to the successors of dest, leaving both
 735  * successors of source NULL.
 736  */
 737
 738 static void
 739 move_successors(nir_block *source, nir_block *dest)
 740 {
 741    nir_block *succ1 = source->successors[0];
 742    nir_block *succ2 = source->successors[1];
 743
 744    if (succ1)
 745       unlink_blocks(source, succ1);
 746
 747    if (succ2)
 748       unlink_blocks(source, succ2);
 749
 750    unlink_block_successors(dest);
 751    link_blocks(dest, succ1, succ2);
 752 }
 753
 754 static nir_block *
 755 split_block_end(nir_block *block)
 756 {
 757    nir_block *new_block = nir_block_create(ralloc_parent(block));
 758    new_block->cf_node.parent = block->cf_node.parent;
 759    exec_node_insert_after(&block->cf_node.node, &new_block->cf_node.node);
 760
 761    move_successors(block, new_block);
 762
 763    return new_block;
 764 }
 765
 766 /**
 767  * Inserts a non-basic block between two basic blocks and links them together.
 768  */
 769
 770 static void
 771 insert_non_block(nir_block *before, nir_cf_node *node, nir_block *after)
 772 {
 773    node->parent = before->cf_node.parent;
 774    exec_node_insert_after(&before->cf_node.node, &node->node);
 775    link_block_to_non_block(before, node);
 776    link_non_block_to_block(node, after);
 777 }
 778
 779 /**
 780  * Inserts a non-basic block before a basic block.
 781  */
 782
 783 static void
 784 insert_non_block_before_block(nir_cf_node *node, nir_block *block)
 785 {
 786    /* split off the beginning of block into new_block */
 787    nir_block *new_block = split_block_beginning(block);
 788
 789    /* insert our node in between new_block and block */
 790    insert_non_block(new_block, node, block);
 791 }
 792
 793 static void
 794 insert_non_block_after_block(nir_block *block, nir_cf_node *node)
 795 {
 796    /* split off the end of block into new_block */
 797    nir_block *new_block = split_block_end(block);
 798
 799    /* insert our node in between block and new_block */
 800    insert_non_block(block, node, new_block);
 801 }
 802
 803 /* walk up the control flow tree to find the innermost enclosed loop */
 804 static nir_loop *
 805 nearest_loop(nir_cf_node *node)
 806 {
 807    while (node->type != nir_cf_node_loop) {
 808       node = node->parent;
 809    }
 810
 811    return nir_cf_node_as_loop(node);
 812 }
 813
 814 nir_function_impl *
 815 nir_cf_node_get_function(nir_cf_node *node)
 816 {
 817    while (node->type != nir_cf_node_function) {
 818       node = node->parent;
 819    }
 820
 821    return nir_cf_node_as_function(node);
 822 }
 823
 824 /*
 825  * update the CFG after a jump instruction has been added to the end of a block
 826  */
 827
 828 static void
 829 handle_jump(nir_block *block)
 830 {
 831    nir_instr *instr = nir_block_last_instr(block);
 832    nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
 833
 834    unlink_block_successors(block);
 835
 836    nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
 837    nir_metadata_dirty(impl, nir_metadata_none);
 838
 839    if (jump_instr->type == nir_jump_break ||
 840        jump_instr->type == nir_jump_continue) {
 841       nir_loop *loop = nearest_loop(&block->cf_node);
 842
 843       if (jump_instr->type == nir_jump_continue) {
 844          nir_cf_node *first_node = nir_loop_first_cf_node(loop);
 845          assert(first_node->type == nir_cf_node_block);
 846          nir_block *first_block = nir_cf_node_as_block(first_node);
 847          link_blocks(block, first_block, NULL);
 848       } else {
 849          nir_cf_node *after = nir_cf_node_next(&loop->cf_node);
 850          assert(after->type == nir_cf_node_block);
 851          nir_block *after_block = nir_cf_node_as_block(after);
 852          link_blocks(block, after_block, NULL);
 853
 854          /* If we inserted a fake link, remove it */
 855          nir_cf_node *last = nir_loop_last_cf_node(loop);
 856          assert(last->type == nir_cf_node_block);
 857          nir_block *last_block =  nir_cf_node_as_block(last);
 858          if (last_block->successors[1] != NULL)
 859             unlink_blocks(last_block, after_block);
 860       }
 861    } else {
 862       assert(jump_instr->type == nir_jump_return);
 863       link_blocks(block, impl->end_block, NULL);
 864    }
 865 }
 866
 867 static void
 868 handle_remove_jump(nir_block *block, nir_jump_type type)
 869 {
 870    unlink_block_successors(block);
 871
 872    if (exec_node_is_tail_sentinel(block->cf_node.node.next)) {
 873       nir_cf_node *parent = block->cf_node.parent;
 874       if (parent->type == nir_cf_node_if) {
 875          nir_cf_node *next = nir_cf_node_next(parent);
 876          assert(next->type == nir_cf_node_block);
 877          nir_block *next_block = nir_cf_node_as_block(next);
 878
 879          link_blocks(block, next_block, NULL);
 880       } else {
 881          assert(parent->type == nir_cf_node_loop);
 882          nir_loop *loop = nir_cf_node_as_loop(parent);
 883
 884          nir_cf_node *head = nir_loop_first_cf_node(loop);
 885          assert(head->type == nir_cf_node_block);
 886          nir_block *head_block = nir_cf_node_as_block(head);
 887
 888          link_blocks(block, head_block, NULL);
 889       }
 890    } else {
 891       nir_cf_node *next = nir_cf_node_next(&block->cf_node);
 892       if (next->type == nir_cf_node_if) {
 893          nir_if *next_if = nir_cf_node_as_if(next);
 894
 895          nir_cf_node *first_then = nir_if_first_then_node(next_if);
 896          assert(first_then->type == nir_cf_node_block);
 897          nir_block *first_then_block = nir_cf_node_as_block(first_then);
 898
 899          nir_cf_node *first_else = nir_if_first_else_node(next_if);
 900          assert(first_else->type == nir_cf_node_block);
 901          nir_block *first_else_block = nir_cf_node_as_block(first_else);
 902
 903          link_blocks(block, first_then_block, first_else_block);
 904       } else {
 905          assert(next->type == nir_cf_node_loop);
 906          nir_loop *next_loop = nir_cf_node_as_loop(next);
 907
 908          nir_cf_node *first = nir_loop_first_cf_node(next_loop);
 909          assert(first->type == nir_cf_node_block);
 910          nir_block *first_block = nir_cf_node_as_block(first);
 911
 912          link_blocks(block, first_block, NULL);
 913       }
 914    }
 915
 916    if (type == nir_jump_break) {
 917       nir_loop *loop = nearest_loop(&block->cf_node);
 918
 919       nir_cf_node *next = nir_cf_node_next(&loop->cf_node);
 920       assert(next->type == nir_cf_node_block);
 921       nir_block *next_block = nir_cf_node_as_block(next);
 922
 923       if (next_block->predecessors->entries == 0) {
 924          /* insert fake link */
 925          nir_cf_node *last = nir_loop_last_cf_node(loop);
 926          assert(last->type == nir_cf_node_block);
 927          nir_block *last_block = nir_cf_node_as_block(last);
 928
 929          last_block->successors[1] = next_block;
 930          block_add_pred(next_block, last_block);
 931       }
 932    }
 933
 934    nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
 935    nir_metadata_dirty(impl, nir_metadata_none);
 936 }
 937
 938 /**
 939  * Inserts a basic block before another by merging the instructions.
 940  *
 941  * @param block the target of the insertion
 942  * @param before the block to be inserted - must not have been inserted before
 943  * @param has_jump whether \before has a jump instruction at the end
 944  */
 945
 946 static void
 947 insert_block_before_block(nir_block *block, nir_block *before, bool has_jump)
 948 {
 949    assert(!has_jump || exec_list_is_empty(&block->instr_list));
 950
 951    foreach_list_typed(nir_instr, instr, node, &before->instr_list) {
 952       instr->block = block;
 953    }
 954
 955    exec_list_prepend(&block->instr_list, &before->instr_list);
 956
 957    if (has_jump)
 958       handle_jump(block);
 959 }
 960
 961 /**
 962  * Inserts a basic block after another by merging the instructions.
 963  *
 964  * @param block the target of the insertion
 965  * @param after the block to be inserted - must not have been inserted before
 966  * @param has_jump whether \after has a jump instruction at the end
 967  */
 968
 969 static void
 970 insert_block_after_block(nir_block *block, nir_block *after, bool has_jump)
 971 {
 972    foreach_list_typed(nir_instr, instr, node, &after->instr_list) {
 973       instr->block = block;
 974    }
 975
 976    exec_list_append(&block->instr_list, &after->instr_list);
 977
 978    if (has_jump)
 979       handle_jump(block);
 980 }
 981
 982 static void
 983 update_if_uses(nir_cf_node *node)
 984 {
 985    if (node->type != nir_cf_node_if)
 986       return;
 987
 988    nir_if *if_stmt = nir_cf_node_as_if(node);
 989    if (if_stmt->condition.is_ssa)
 990       return;
 991
 992    nir_register *reg = if_stmt->condition.reg.reg;
 993    assert(reg != NULL);
 994
 995    _mesa_set_add(reg->if_uses, _mesa_hash_pointer(if_stmt), if_stmt);
 996 }
 997
 998 void
 999 nir_cf_node_insert_after(nir_cf_node *node, nir_cf_node *after)
1000 {
1001    update_if_uses(after);
1002
1003    if (after->type == nir_cf_node_block) {
1004       /*
1005        * either node or the one after it must be a basic block, by invariant #2;
1006        * in either case, just merge the blocks together.
1007        */
1008       nir_block *after_block = nir_cf_node_as_block(after);
1009
1010       bool has_jump = !exec_list_is_empty(&after_block->instr_list) &&
1011          nir_block_last_instr(after_block)->type == nir_instr_type_jump;
1012
1013       if (node->type == nir_cf_node_block) {
1014          insert_block_after_block(nir_cf_node_as_block(node), after_block,
1015                                   has_jump);
1016       } else {
1017          nir_cf_node *next = nir_cf_node_next(node);
1018          assert(next->type == nir_cf_node_block);
1019          nir_block *next_block = nir_cf_node_as_block(next);
1020
1021          insert_block_before_block(next_block, after_block, has_jump);
1022       }
1023    } else {
1024       if (node->type == nir_cf_node_block) {
1025          insert_non_block_after_block(nir_cf_node_as_block(node), after);
1026       } else {
1027          /*
1028           * We have to insert a non-basic block after a non-basic block. Since
1029           * every non-basic block has a basic block after it, this is equivalent
1030           * to inserting a non-basic block before a basic block.
1031           */
1032
1033          nir_cf_node *next = nir_cf_node_next(node);
1034          assert(next->type == nir_cf_node_block);
1035          nir_block *next_block = nir_cf_node_as_block(next);
1036
1037          insert_non_block_before_block(after, next_block);
1038       }
1039    }
1040
1041    nir_function_impl *impl = nir_cf_node_get_function(node);
1042    nir_metadata_dirty(impl, nir_metadata_none);
1043 }
1044
1045 void
1046 nir_cf_node_insert_before(nir_cf_node *node, nir_cf_node *before)
1047 {
1048    update_if_uses(before);
1049
1050    if (before->type == nir_cf_node_block) {
1051       nir_block *before_block = nir_cf_node_as_block(before);
1052
1053       bool has_jump = !exec_list_is_empty(&before_block->instr_list) &&
1054          nir_block_last_instr(before_block)->type == nir_instr_type_jump;
1055
1056       if (node->type == nir_cf_node_block) {
1057          insert_block_before_block(nir_cf_node_as_block(node), before_block,
1058                                    has_jump);
1059       } else {
1060          nir_cf_node *prev = nir_cf_node_prev(node);
1061          assert(prev->type == nir_cf_node_block);
1062          nir_block *prev_block = nir_cf_node_as_block(prev);
1063
1064          insert_block_after_block(prev_block, before_block, has_jump);
1065       }
1066    } else {
1067       if (node->type == nir_cf_node_block) {
1068          insert_non_block_before_block(before, nir_cf_node_as_block(node));
1069       } else {
1070          /*
1071           * We have to insert a non-basic block before a non-basic block. This
1072           * is equivalent to inserting a non-basic block after a basic block.
1073           */
1074
1075          nir_cf_node *prev_node = nir_cf_node_prev(node);
1076          assert(prev_node->type == nir_cf_node_block);
1077          nir_block *prev_block = nir_cf_node_as_block(prev_node);
1078
1079          insert_non_block_after_block(prev_block, before);
1080       }
1081    }
1082
1083    nir_function_impl *impl = nir_cf_node_get_function(node);
1084    nir_metadata_dirty(impl, nir_metadata_none);
1085 }
1086
1087 void
1088 nir_cf_node_insert_begin(struct exec_list *list, nir_cf_node *node)
1089 {
1090    nir_cf_node *begin = exec_node_data(nir_cf_node, list->head, node);
1091    nir_cf_node_insert_before(begin, node);
1092 }
1093
1094 void
1095 nir_cf_node_insert_end(struct exec_list *list, nir_cf_node *node)
1096 {
1097    nir_cf_node *end = exec_node_data(nir_cf_node, list->tail_pred, node);
1098    nir_cf_node_insert_after(end, node);
1099 }
1100
1101 /**
1102  * Stitch two basic blocks together into one. The aggregate must have the same
1103  * predecessors as the first and the same successors as the second.
1104  */
1105
1106 static void
1107 stitch_blocks(nir_block *before, nir_block *after)
1108 {
1109    /*
1110     * We move after into before, so we have to deal with up to 2 successors vs.
1111     * possibly a large number of predecessors.
1112     *
1113     * TODO: special case when before is empty and after isn't?
1114     */
1115
1116    move_successors(after, before);
1117
1118    foreach_list_typed(nir_instr, instr, node, &after->instr_list) {
1119       instr->block = before;
1120    }
1121
1122    exec_list_append(&before->instr_list, &after->instr_list);
1123    exec_node_remove(&after->cf_node.node);
1124 }
1125
1126 void
1127 nir_cf_node_remove(nir_cf_node *node)
1128 {
1129    nir_function_impl *impl = nir_cf_node_get_function(node);
1130    nir_metadata_dirty(impl, nir_metadata_none);
1131
1132    if (node->type == nir_cf_node_block) {
1133       /*
1134        * Basic blocks can't really be removed by themselves, since they act as
1135        * padding between the non-basic blocks. So all we do here is empty the
1136        * block of instructions.
1137        *
1138        * TODO: could we assert here?
1139        */
1140       exec_list_make_empty(&nir_cf_node_as_block(node)->instr_list);
1141    } else {
1142       nir_cf_node *before = nir_cf_node_prev(node);
1143       assert(before->type == nir_cf_node_block);
1144       nir_block *before_block = nir_cf_node_as_block(before);
1145
1146       nir_cf_node *after = nir_cf_node_next(node);
1147       assert(after->type == nir_cf_node_block);
1148       nir_block *after_block = nir_cf_node_as_block(after);
1149
1150       exec_node_remove(&node->node);
1151       stitch_blocks(before_block, after_block);
1152    }
1153 }
1154
1155 static bool
1156 add_use_cb(nir_src *src, void *state)
1157 {
1158    nir_instr *instr = (nir_instr *) state;
1159
1160    struct set *uses_set = src->is_ssa ? src->ssa->uses : src->reg.reg->uses;
1161
1162    _mesa_set_add(uses_set, _mesa_hash_pointer(instr), instr);
1163
1164    return true;
1165 }
1166
1167 static bool
1168 add_def_cb(nir_dest *dest, void *state)
1169 {
1170    nir_instr *instr = (nir_instr *) state;
1171
1172    if (dest->is_ssa)
1173       return true;
1174
1175    nir_register *reg = dest->reg.reg;
1176
1177    _mesa_set_add(reg->defs, _mesa_hash_pointer(instr), instr);
1178
1179    return true;
1180 }
1181
1182 static void
1183 add_defs_uses(nir_instr *instr)
1184 {
1185    nir_foreach_src(instr, add_use_cb, instr);
1186    nir_foreach_dest(instr, add_def_cb, instr);
1187 }
1188
1189 void
1190 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
1191 {
1192    assert(before->type != nir_instr_type_jump);
1193    before->block = instr->block;
1194    add_defs_uses(before);
1195    exec_node_insert_node_before(&instr->node, &before->node);
1196 }
1197
1198 void
1199 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
1200 {
1201    if (after->type == nir_instr_type_jump) {
1202       assert(instr == nir_block_last_instr(instr->block));
1203       assert(instr->type != nir_instr_type_jump);
1204    }
1205
1206    after->block = instr->block;
1207    add_defs_uses(after);
1208    exec_node_insert_after(&instr->node, &after->node);
1209
1210    if (after->type == nir_instr_type_jump)
1211       handle_jump(after->block);
1212 }
1213
1214 void
1215 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
1216 {
1217    if (before->type == nir_instr_type_jump)
1218       assert(exec_list_is_empty(&block->instr_list));
1219
1220    before->block = block;
1221    add_defs_uses(before);
1222    exec_list_push_head(&block->instr_list, &before->node);
1223
1224    if (before->type == nir_instr_type_jump)
1225       handle_jump(block);
1226 }
1227
1228 void
1229 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
1230 {
1231    if (after->type == nir_instr_type_jump) {
1232       assert(exec_list_is_empty(&block->instr_list) ||
1233              nir_block_last_instr(block)->type != nir_instr_type_jump);
1234    }
1235
1236    after->block = block;
1237    add_defs_uses(after);
1238    exec_list_push_tail(&block->instr_list, &after->node);
1239
1240    if (after->type == nir_instr_type_jump)
1241       handle_jump(block);
1242 }
1243
1244 void
1245 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
1246 {
1247    if (node->type == nir_cf_node_block) {
1248       nir_instr_insert_before_block(nir_cf_node_as_block(node), before);
1249    } else {
1250       nir_cf_node *prev = nir_cf_node_prev(node);
1251       assert(prev->type == nir_cf_node_block);
1252       nir_block *prev_block = nir_cf_node_as_block(prev);
1253
1254       nir_instr_insert_before_block(prev_block, before);
1255    }
1256 }
1257
1258 void
1259 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
1260 {
1261       if (node->type == nir_cf_node_block) {
1262       nir_instr_insert_after_block(nir_cf_node_as_block(node), after);
1263    } else {
1264       nir_cf_node *next = nir_cf_node_next(node);
1265       assert(next->type == nir_cf_node_block);
1266       nir_block *next_block = nir_cf_node_as_block(next);
1267
1268       nir_instr_insert_before_block(next_block, after);
1269    }
1270 }
1271
1272 void
1273 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
1274 {
1275    nir_cf_node *first_node = exec_node_data(nir_cf_node,
1276                                             exec_list_get_head(list), node);
1277    nir_instr_insert_before_cf(first_node, before);
1278 }
1279
1280 void
1281 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
1282 {
1283    nir_cf_node *last_node = exec_node_data(nir_cf_node,
1284                                            exec_list_get_tail(list), node);
1285    nir_instr_insert_after_cf(last_node, after);
1286 }
1287
1288 static bool
1289 remove_use_cb(nir_src *src, void *state)
1290 {
1291    nir_instr *instr = (nir_instr *) state;
1292
1293    struct set *uses_set = src->is_ssa ? src->ssa->uses : src->reg.reg->uses;
1294
1295    struct set_entry *entry = _mesa_set_search(uses_set,
1296                                               _mesa_hash_pointer(instr),
1297                                               instr);
1298    if (entry)
1299       _mesa_set_remove(uses_set, entry);
1300
1301    return true;
1302 }
1303
1304 static bool
1305 remove_def_cb(nir_dest *dest, void *state)
1306 {
1307    nir_instr *instr = (nir_instr *) state;
1308
1309    if (dest->is_ssa)
1310       return true;
1311
1312    nir_register *reg = dest->reg.reg;
1313
1314    struct set_entry *entry = _mesa_set_search(reg->defs,
1315                                               _mesa_hash_pointer(instr),
1316                                               instr);
1317    if (entry)
1318       _mesa_set_remove(reg->defs, entry);
1319
1320    return true;
1321 }
1322
1323 static void
1324 remove_defs_uses(nir_instr *instr)
1325 {
1326    nir_foreach_dest(instr, remove_def_cb, instr);
1327    nir_foreach_src(instr, remove_use_cb, instr);
1328 }
1329
1330 void nir_instr_remove(nir_instr *instr)
1331 {
1332    remove_defs_uses(instr);
1333    exec_node_remove(&instr->node);
1334
1335    if (instr->type == nir_instr_type_jump) {
1336       nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
1337       handle_remove_jump(instr->block, jump_instr->type);
1338    }
1339 }
1340
1341 /*@}*/
1342
1343 void
1344 nir_index_local_regs(nir_function_impl *impl)
1345 {
1346    unsigned index = 0;
1347    foreach_list_typed(nir_register, reg, node, &impl->registers) {
1348       reg->index = index++;
1349    }
1350    impl->reg_alloc = index;
1351 }
1352
1353 void
1354 nir_index_global_regs(nir_shader *shader)
1355 {
1356    unsigned index = 0;
1357    foreach_list_typed(nir_register, reg, node, &shader->registers) {
1358       reg->index = index++;
1359    }
1360    shader->reg_alloc = index;
1361 }
1362
1363 static bool
1364 visit_alu_dest(nir_alu_instr *instr, nir_foreach_dest_cb cb, void *state)
1365 {
1366    return cb(&instr->dest.dest, state);
1367 }
1368
1369 static bool
1370 visit_intrinsic_dest(nir_intrinsic_instr *instr, nir_foreach_dest_cb cb,
1371                      void *state)
1372 {
1373    if (nir_intrinsic_infos[instr->intrinsic].has_dest)
1374       return cb(&instr->dest, state);
1375
1376    return true;
1377 }
1378
1379 static bool
1380 visit_texture_dest(nir_tex_instr *instr, nir_foreach_dest_cb cb,
1381                    void *state)
1382 {
1383    return cb(&instr->dest, state);
1384 }
1385
1386 static bool
1387 visit_load_const_dest(nir_load_const_instr *instr, nir_foreach_dest_cb cb,
1388                       void *state)
1389 {
1390    return cb(&instr->dest, state);
1391 }
1392
1393 static bool
1394 visit_phi_dest(nir_phi_instr *instr, nir_foreach_dest_cb cb, void *state)
1395 {
1396    return cb(&instr->dest, state);
1397 }
1398
1399 static bool
1400 visit_parallel_copy_dest(nir_parallel_copy_instr *instr,
1401                          nir_foreach_dest_cb cb, void *state)
1402 {
1403    foreach_list_typed(nir_parallel_copy_copy, copy, node, &instr->copies) {
1404       if (!cb(&copy->dest, state))
1405          return false;
1406    }
1407
1408    return true;
1409 }
1410
1411 bool
1412 nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state)
1413 {
1414    switch (instr->type) {
1415    case nir_instr_type_alu:
1416       return visit_alu_dest(nir_instr_as_alu(instr), cb, state);
1417    case nir_instr_type_intrinsic:
1418       return visit_intrinsic_dest(nir_instr_as_intrinsic(instr), cb, state);
1419    case nir_instr_type_texture:
1420       return visit_texture_dest(nir_instr_as_texture(instr), cb, state);
1421    case nir_instr_type_load_const:
1422       return visit_load_const_dest(nir_instr_as_load_const(instr), cb, state);
1423    case nir_instr_type_phi:
1424       return visit_phi_dest(nir_instr_as_phi(instr), cb, state);
1425    case nir_instr_type_parallel_copy:
1426       return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr),
1427                                       cb, state);
1428
1429    case nir_instr_type_ssa_undef:
1430    case nir_instr_type_call:
1431    case nir_instr_type_jump:
1432       break;
1433
1434    default:
1435       unreachable("Invalid instruction type");
1436       break;
1437    }
1438
1439    return true;
1440 }
1441
1442 static bool
1443 visit_src(nir_src *src, nir_foreach_src_cb cb, void *state)
1444 {
1445    if (!cb(src, state))
1446       return false;
1447    if (!src->is_ssa && src->reg.indirect)
1448       return cb(src->reg.indirect, state);
1449    return true;
1450 }
1451
1452 static bool
1453 visit_deref_array_src(nir_deref_array *deref, nir_foreach_src_cb cb,
1454                       void *state)
1455 {
1456    if (deref->has_indirect)
1457       return visit_src(&deref->indirect, cb, state);
1458    return true;
1459 }
1460
1461 static bool
1462 visit_deref_src(nir_deref_var *deref, nir_foreach_src_cb cb, void *state)
1463 {
1464    nir_deref *cur = &deref->deref;
1465    while (cur != NULL) {
1466       if (cur->deref_type == nir_deref_type_array)
1467          if (!visit_deref_array_src(nir_deref_as_array(cur), cb, state))
1468             return false;
1469
1470       cur = cur->child;
1471    }
1472
1473    return true;
1474 }
1475
1476 static bool
1477 visit_alu_src(nir_alu_instr *instr, nir_foreach_src_cb cb, void *state)
1478 {
1479    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1480       if (!visit_src(&instr->src[i].src, cb, state))
1481          return false;
1482
1483    if (instr->has_predicate)
1484       if (!visit_src(&instr->predicate, cb, state))
1485          return false;
1486
1487    return true;
1488 }
1489
1490 static bool
1491 visit_tex_src(nir_tex_instr *instr, nir_foreach_src_cb cb, void *state)
1492 {
1493    for (unsigned i = 0; i < instr->num_srcs; i++)
1494       if (!visit_src(&instr->src[i], cb, state))
1495          return false;
1496
1497    if (instr->has_predicate)
1498       if (!visit_src(&instr->predicate, cb, state))
1499          return false;
1500
1501    if (instr->sampler != NULL)
1502       if (!visit_deref_src(instr->sampler, cb, state))
1503          return false;
1504
1505    return true;
1506 }
1507
1508 static bool
1509 visit_intrinsic_src(nir_intrinsic_instr *instr, nir_foreach_src_cb cb,
1510                     void *state)
1511 {
1512    unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs;
1513    for (unsigned i = 0; i < num_srcs; i++)
1514       if (!visit_src(&instr->src[i], cb, state))
1515          return false;
1516
1517    unsigned num_vars =
1518       nir_intrinsic_infos[instr->intrinsic].num_variables;
1519    for (unsigned i = 0; i < num_vars; i++)
1520       if (!visit_deref_src(instr->variables[i], cb, state))
1521          return false;
1522
1523    if (instr->has_predicate)
1524       if (!visit_src(&instr->predicate, cb, state))
1525          return false;
1526
1527    return true;
1528 }
1529
1530 static bool
1531 visit_call_src(nir_call_instr *instr, nir_foreach_src_cb cb, void *state)
1532 {
1533    if (instr->has_predicate)
1534       if (!visit_src(&instr->predicate, cb, state))
1535          return false;
1536
1537    return true;
1538 }
1539
1540 static bool
1541 visit_load_const_src(nir_load_const_instr *instr, nir_foreach_src_cb cb,
1542                      void *state)
1543 {
1544    if (instr->has_predicate)
1545       if (!visit_src(&instr->predicate, cb, state))
1546          return false;
1547
1548    return true;
1549 }
1550
1551 static bool
1552 visit_phi_src(nir_phi_instr *instr, nir_foreach_src_cb cb, void *state)
1553 {
1554    foreach_list_typed(nir_phi_src, src, node, &instr->srcs) {
1555       if (!visit_src(&src->src, cb, state))
1556          return false;
1557    }
1558
1559    return true;
1560 }
1561
1562 static bool
1563 visit_parallel_copy_src(nir_parallel_copy_instr *instr,
1564                         nir_foreach_src_cb cb, void *state)
1565 {
1566    foreach_list_typed(nir_parallel_copy_copy, copy, node, &instr->copies) {
1567       if (!visit_src(&copy->src, cb, state))
1568          return false;
1569    }
1570
1571    return true;
1572 }
1573
1574 typedef struct {
1575    void *state;
1576    nir_foreach_src_cb cb;
1577 } visit_dest_indirect_state;
1578
1579 static bool
1580 visit_dest_indirect(nir_dest *dest, void *_state)
1581 {
1582    visit_dest_indirect_state *state = (visit_dest_indirect_state *) _state;
1583
1584    if (!dest->is_ssa && dest->reg.indirect)
1585       return state->cb(dest->reg.indirect, state->state);
1586
1587    return true;
1588 }
1589
1590 bool
1591 nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state)
1592 {
1593    switch (instr->type) {
1594    case nir_instr_type_alu:
1595       if (!visit_alu_src(nir_instr_as_alu(instr), cb, state))
1596          return false;
1597       break;
1598    case nir_instr_type_intrinsic:
1599       if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr), cb, state))
1600          return false;
1601       break;
1602    case nir_instr_type_texture:
1603       if (!visit_tex_src(nir_instr_as_texture(instr), cb, state))
1604          return false;
1605       break;
1606    case nir_instr_type_call:
1607       if (!visit_call_src(nir_instr_as_call(instr), cb, state))
1608          return false;
1609       break;
1610    case nir_instr_type_load_const:
1611       if (!visit_load_const_src(nir_instr_as_load_const(instr), cb, state))
1612          return false;
1613       break;
1614    case nir_instr_type_phi:
1615       if (!visit_phi_src(nir_instr_as_phi(instr), cb, state))
1616          return false;
1617       break;
1618    case nir_instr_type_parallel_copy:
1619       if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr),
1620                                    cb, state))
1621          return false;
1622       break;
1623    case nir_instr_type_jump:
1624    case nir_instr_type_ssa_undef:
1625       return true;
1626
1627    default:
1628       unreachable("Invalid instruction type");
1629       break;
1630    }
1631
1632    visit_dest_indirect_state dest_state;
1633    dest_state.state = state;
1634    dest_state.cb = cb;
1635    return nir_foreach_dest(instr, visit_dest_indirect, &dest_state);
1636 }
1637
1638 bool
1639 nir_srcs_equal(nir_src src1, nir_src src2)
1640 {
1641    if (src1.is_ssa) {
1642       if (src2.is_ssa) {
1643          return src1.ssa == src2.ssa;
1644       } else {
1645          return false;
1646       }
1647    } else {
1648       if (src2.is_ssa) {
1649          return false;
1650       } else {
1651          if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
1652             return false;
1653
1654          if (src1.reg.indirect) {
1655             if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
1656                return false;
1657          }
1658
1659          return src1.reg.reg == src2.reg.reg &&
1660                 src1.reg.base_offset == src2.reg.base_offset;
1661       }
1662    }
1663 }
1664
1665 void
1666 nir_ssa_def_init(nir_function_impl *impl, nir_instr *instr, nir_ssa_def *def,
1667                  unsigned num_components, const char *name)
1668 {
1669    void *mem_ctx = ralloc_parent(instr);
1670
1671    def->name = name;
1672    def->index = impl->ssa_alloc++;
1673    def->parent_instr = instr;
1674    def->uses = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
1675    def->if_uses = _mesa_set_create(mem_ctx, _mesa_key_pointer_equal);
1676    def->num_components = num_components;
1677 }
1678
1679 struct ssa_def_rewrite_state {
1680    void *mem_ctx;
1681    nir_ssa_def *old;
1682    nir_src new_src;
1683 };
1684
1685 static bool
1686 ssa_def_rewrite_uses_src(nir_src *src, void *void_state)
1687 {
1688    struct ssa_def_rewrite_state *state = void_state;
1689
1690    if (src->is_ssa && src->ssa == state->old)
1691       *src = nir_src_copy(state->new_src, state->mem_ctx);
1692
1693    return true;
1694 }
1695
1696 void
1697 nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src, void *mem_ctx)
1698 {
1699    struct ssa_def_rewrite_state state;
1700    state.mem_ctx = mem_ctx;
1701    state.old = def;
1702    state.new_src = new_src;
1703
1704    assert(!new_src.is_ssa || def != new_src.ssa);
1705
1706    struct set *new_uses, *new_if_uses;
1707    if (new_src.is_ssa) {
1708       new_uses = new_src.ssa->uses;
1709       new_if_uses = new_src.ssa->if_uses;
1710    } else {
1711       new_uses = new_src.reg.reg->uses;
1712       new_if_uses = new_src.reg.reg->if_uses;
1713    }
1714
1715    struct set_entry *entry;
1716    set_foreach(def->uses, entry) {
1717       nir_instr *instr = (nir_instr *)entry->key;
1718
1719       _mesa_set_remove(def->uses, entry);
1720       nir_foreach_src(instr, ssa_def_rewrite_uses_src, &state);
1721       _mesa_set_add(new_uses, _mesa_hash_pointer(instr), instr);
1722    }
1723
1724    set_foreach(def->if_uses, entry) {
1725       nir_if *if_use = (nir_if *)entry->key;
1726
1727       _mesa_set_remove(def->if_uses, entry);
1728       if_use->condition = nir_src_copy(new_src, mem_ctx);
1729       _mesa_set_add(new_if_uses, _mesa_hash_pointer(if_use), if_use);
1730    }
1731 }
1732
1733
1734 static bool foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1735                             bool reverse, void *state);
1736
1737 static inline bool
1738 foreach_if(nir_if *if_stmt, nir_foreach_block_cb cb, bool reverse, void *state)
1739 {
1740    if (reverse) {
1741       foreach_list_typed_reverse(nir_cf_node, node, node, &if_stmt->else_list) {
1742          if (!foreach_cf_node(node, cb, reverse, state))
1743             return false;
1744       }
1745
1746       foreach_list_typed_reverse(nir_cf_node, node, node, &if_stmt->then_list) {
1747          if (!foreach_cf_node(node, cb, reverse, state))
1748             return false;
1749       }
1750    } else {
1751       foreach_list_typed(nir_cf_node, node, node, &if_stmt->then_list) {
1752          if (!foreach_cf_node(node, cb, reverse, state))
1753             return false;
1754       }
1755
1756       foreach_list_typed(nir_cf_node, node, node, &if_stmt->else_list) {
1757          if (!foreach_cf_node(node, cb, reverse, state))
1758             return false;
1759       }
1760    }
1761
1762    return true;
1763 }
1764
1765 static inline bool
1766 foreach_loop(nir_loop *loop, nir_foreach_block_cb cb, bool reverse, void *state)
1767 {
1768    if (reverse) {
1769       foreach_list_typed_reverse(nir_cf_node, node, node, &loop->body) {
1770          if (!foreach_cf_node(node, cb, reverse, state))
1771             return false;
1772       }
1773    } else {
1774       foreach_list_typed(nir_cf_node, node, node, &loop->body) {
1775          if (!foreach_cf_node(node, cb, reverse, state))
1776             return false;
1777       }
1778    }
1779
1780    return true;
1781 }
1782
1783 static bool
1784 foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb,
1785                 bool reverse, void *state)
1786 {
1787    switch (node->type) {
1788    case nir_cf_node_block:
1789       return cb(nir_cf_node_as_block(node), state);
1790    case nir_cf_node_if:
1791       return foreach_if(nir_cf_node_as_if(node), cb, reverse, state);
1792    case nir_cf_node_loop:
1793       return foreach_loop(nir_cf_node_as_loop(node), cb, reverse, state);
1794       break;
1795
1796    default:
1797       unreachable("Invalid CFG node type");
1798       break;
1799    }
1800
1801    return false;
1802 }
1803
1804 bool
1805 nir_foreach_block(nir_function_impl *impl, nir_foreach_block_cb cb, void *state)
1806 {
1807    foreach_list_typed(nir_cf_node, node, node, &impl->body) {
1808       if (!foreach_cf_node(node, cb, false, state))
1809          return false;
1810    }
1811
1812    return cb(impl->end_block, state);
1813 }
1814
1815 bool
1816 nir_foreach_block_reverse(nir_function_impl *impl, nir_foreach_block_cb cb,
1817                           void *state)
1818 {
1819    if (!cb(impl->end_block, state))
1820       return false;
1821
1822    foreach_list_typed_reverse(nir_cf_node, node, node, &impl->body) {
1823       if (!foreach_cf_node(node, cb, true, state))
1824          return false;
1825    }
1826
1827    return true;
1828 }
1829
1830 nir_if *
1831 nir_block_following_if(nir_block *block)
1832 {
1833    if (exec_node_is_tail_sentinel(&block->cf_node.node))
1834       return NULL;
1835
1836    if (nir_cf_node_is_last(&block->cf_node))
1837       return NULL;
1838
1839    nir_cf_node *next_node = nir_cf_node_next(&block->cf_node);
1840
1841    if (next_node->type != nir_cf_node_if)
1842       return NULL;
1843
1844    return nir_cf_node_as_if(next_node);
1845 }
1846
1847 static bool
1848 index_block(nir_block *block, void *state)
1849 {
1850    unsigned *index = (unsigned *) state;
1851    block->index = (*index)++;
1852    return true;
1853 }
1854
1855 void
1856 nir_index_blocks(nir_function_impl *impl)
1857 {
1858    unsigned index = 0;
1859
1860    if (impl->valid_metadata & nir_metadata_block_index)
1861       return;
1862
1863    nir_foreach_block(impl, index_block, &index);
1864
1865    impl->num_blocks = index;
1866 }
1867
1868 static void
1869 index_ssa_def(nir_ssa_def *def, unsigned *index)
1870 {
1871    def->index = (*index)++;
1872 }
1873
1874 static bool
1875 index_ssa_def_cb(nir_dest *dest, void *state)
1876 {
1877    unsigned *index = (unsigned *) state;
1878    if (dest->is_ssa)
1879       index_ssa_def(&dest->ssa, index);
1880    return true;
1881 }
1882
1883 static void
1884 index_ssa_undef(nir_ssa_undef_instr *instr, unsigned *index)
1885 {
1886    index_ssa_def(&instr->def, index);
1887 }
1888
1889 static bool
1890 index_ssa_block(nir_block *block, void *state)
1891 {
1892    unsigned *index = (unsigned *) state;
1893
1894    nir_foreach_instr(block, instr) {
1895       if (instr->type == nir_instr_type_ssa_undef)
1896          index_ssa_undef(nir_instr_as_ssa_undef(instr), index);
1897       else
1898          nir_foreach_dest(instr, index_ssa_def_cb, state);
1899    }
1900
1901    return true;
1902 }
1903
1904 void
1905 nir_index_ssa_defs(nir_function_impl *impl)
1906 {
1907    unsigned index = 0;
1908    nir_foreach_block(impl, index_ssa_block, &index);
1909    impl->ssa_alloc = index;
1910 }